did the thing. still experimentali suggest using 12 steps with the turbo lora i use.further testing on civitai generator, the difference in quality is insane. i really do suggest using comfyui 4mp for best results for the hi res lora. the v1.0 version does well in low resolutions.The hires lora can be used up to 1.50 strength or probably more. The images i did with fineview checkpoint model, i used 1.50New: mrpopo hires moss further: I suggest using euler or gradient_estimation for a bit extra detail. I found that er_sde dosen't work nicely for me in my test.Just better overall than last model. I highly suggest using Moss base. use with turbo model that i use, lodestone dmd2 base. The strength i suggest for that turbo lora is 1.35. My lora,I just leave it as 1.00 strength. Yes, it makes a big difference. Also, i have to say it again, best used in comfyui or where ever else you can use up to 4mp resolution. This lora is designed to be used in 4mp resolution. However it can still use lower resolutions, just not tested enough to kno if its good. Got my other lora for that though.( v1.0).I personally think this is the most "realistic" lora out there for Moss model. If you like cinematic stuff, this is the one to get. It also does anime/art as well. It just does everything really. As long as you prompt it correctly of course.Note: my lora is meant for base model, it may have issues if used in sft model or loras made for sft.Have fun with it :)

Mrpopo's clarity-moss-image

i retrained the lora using the same dataset, but including new tweaks in the training settings. Was trained on ostris toolkit cloud. gpu used was b200. Trained only on 2048 resolution. yes i know, very costly but worth it. Noticeable changes: just everything looks better, as it should. Better skin details for close up shots.(finally, there were a lot of plastic skin. Still prevalent but better) Lighting got a big upgrade. You know what i say, lighting is what makes realism so i made big emphasis on that. Note: that this lora will completely change moss image model. Thats what i designed for this lora to do. Whether thats a good thing or bad, up to you to decide. I highly recommend using the base model. Sft might work but may be choppy. This is because sft has big bias on lower resolution output, making it look plastic. it fights my hires lora. Therefore, use base with turbo model. i suggest the loadestones dmd2 strength 1.35 is perfect.

Files with SHA256: 51304a696330bb686cbef917a5c2bb70d6d9021a94c10cf1ac34823cd5de9dcd